Why Some AI Agents Start Acting Like Family, and What That Reveals About Cooperation

Ever thought why individuals cooperate, even at personal cost, for many decades? Well, biologists have sought answers to the question. Surprisingly, one of the most widely accepted theories is inclusive fitness theory. It argues that individuals are more likely to cooperate with others who share their genes, as it indirectly contributes to their survival.

As researchers are developing multi-agent reinforcement learning systems based on these biological principles of cooperation, this concept of cooperation is now being extended into the realm of artificial intelligence.


Researchers added a concept of genetic relatedness to reward systems, allowing artificial agents to receive more rewards by assisting others who are similar, argues a study published in 2025 on arXiv.

This creates a pattern of cooperation akin to animal societies. This is because, in conventional reinforcement learning, the agent works towards optimizing its reward, but the addition of inclusive fitness makes the reward function dependent on the performance of other agents.

What does this imply? The agent will work towards cooperation in circumstances where it will benefit from the performance of the other agent.

Evolutionary dynamics and long-term cooperation

The addition of the evolutionary dynamics brings an additional level of complexity. Wondering why? This is because the agents not only learn from the immediate rewards, but they also evolve strategies over time based on the changing environment and interactions.

Analogous to biological evolution, this is where cooperation arises based on the long-term benefits outweighing the short-term costs.

Social networks that evolve on their own

Another major development in the field is the ability for the agents to modify their social connections depending on their experiences, thus improving their cooperative connections. This is achieved by the use of adaptive rewiring strategies, where the agents apply Q-learning to remove connections with non-cooperative agents and form connections with cooperative ones.

A study published in 2025 on the arXiv website indicates that the process leads to the development of cooperative clusters, which operate similarly to animal groups depending on their mutual trust and benefits. The dynamic nature of the network is crucial for the development of cooperative connections, especially in large and complex systems.

In addition to the genetic similarity, scientists are also investigating the effect of the reputation systems on cooperation.

What this means for the future of AI

These results indicate that cooperation does not have to be programmed into artificial systems but can arise as a result of simple rules derived from biology, which is a significant finding with regard to the creation of more efficient artificial intelligence systems.

By using the theories of inclusive fitness, adaptive networks, and reputation systems together, scientists are able to create systems that do not behave like traditional machines but like social organisms.